• 沒有找到結果。

3. PROPOSED DISCRETIZATION MODEL

3.3 Multi-objective Model

From previous section, we can discover that our objectives seem to be conflict between the number of intervals and rules, accuracy, and upper bound value. Therefore, a multi-objective optimization model needs to be added in the discretization process, as in the design Fig. 3-2.

3.3.1 Notations

First, we give out some mathematical notations and definitions of parameters in Table 3-2.

Notice that here we set an upper bound for split points and a lower bound for rules; those bounds are determined by researchers.

Table 3-2 Notations and definitions Terms used in model Notation and definitions

Sum of number of split points 𝑝 ∈ 𝑁 , 𝑝 = ∑ 𝑎𝑙𝑙 𝑠𝑝𝑙𝑖𝑡 𝑝𝑜𝑖𝑛𝑡𝑠

Upper bound of split points 𝑛𝑝 ∈ 𝑁, 𝑝 ≤ 𝑛𝑝 (value is determined by researchers)

Number of rules 𝑟 ∈ 𝑁

Lower bound of rules 𝑛𝑅 ∈ 𝑁, 𝑛𝑅 ≤ 𝑛 (value is determined by researchers)

Accuracy 𝑎, 0 ≤ 𝑎 ≤ 1

Weight of each objective 𝑤𝑖 where 𝑖 = 1,2, … 0 ≤ 𝑤𝑖 ≤ 1, ∑ 𝑤𝑖 = 1

30

Figure 3-2 Proposed Multi-objective Model

31

3.3.2 Fitness function

From the definitions and notations, here we standardize objectives into a number between 0 and 1. Besides, we alternated accuracy into error rate for finding minimum. After standardization, we have four functions and need to weight each function since we choose weight method.

Table 3-3 Standardization

Objective Standardization Target

Number of intervals 𝑓1 = 𝑛𝑝

𝑝, 0 ≤ 𝑓1 ≤ 1 Minimum

Number of rules 𝑓2 =𝑟−𝑛r𝑅, 0 ≤ 𝑓2 ≤ 1 Minimum

Error rate 𝑓3 = 𝜀, 0 ≤ 𝑓3 ≤ 1 Minimum

Fitness function denotes as:

𝐹 = 𝑤1𝑓1+ 𝑤2𝑓2+ 𝑤3𝑓3 Eq. 3-1

where 𝑤1, 𝑤2, 𝑤3 are the weight of the four objective functions

The optimal solution is to find the minimum of the fitness function.

32

3.3.3 Pseudo code

FOR (ITERATION TIMES = 10000) {

FOR (ALL ATTRIBUTES ONE BY ONE) { //discretization SORT;

GENERATE SPLIT POINTS (TOTAL = UPPER BOUND);

FOR (ALL SPLIT POINTS) {

IF (SPLIT POINTS ARE EQUAL) { SPLIT POINTS – 1; }

}

CREATE INTERVALS;

}

FOR (ITERATION TIMES = 1000) { //generate classification rules GENERATE RULES (RANDOM NUMBER UNDER UPPER BOUND);

CALCULATE ACCURACY;

}

CALCULATE FITNESS;

PSO CORRECTING SPILT POINTS POSITION;

}

33

3.3.4 Emulation Picture

Step 1: Randomly select dividing points (we can set up the maximum number of bins)

Figure 3-3 randomly select dividing points

Step 2: List all possible rule set and calculate affinity degree

Figure 3-4 List all possible rule set and calculate affinity degree

34

Step 3: Iterate selecting rules randomly and calculate fitness

Figure 3-5 Iterate selecting rules randomly

Step 4: select dividing bins via PSO

Figure 3-6 select dividing bins via PSO

35

Step 5: Iterate Step 2 to Step 4, until reaching terminating conditions.

Figure 3-7 Final dividing bins

36

4. DATA ANALYSIS

In this paper we use two datasets: the well-known iris flower data set, and multiple enrollment program and academic achievements dataset attained in St. John's University. The details of the datasets are listed as in Table4-1.

Table 4-1 Details of datasets

Data set Examples Attributes Classes

IRIS flower

data set

150 Sepal length (continuous) Sepal width (continuous)

460 Academic achievements average (continuous)

Physical training score (continuous)

Score rank in class (continuous) College department (discrete)

37

IRIS flower data set was obtained from University of California at Irvine (UCI)’s data set repository, and multi enrollment dataset of multi enrollment programs in Taiwan’s technical–

vocational college was obtained from a 4-year system technical–vocational college in Taipei (personal information was removed for protecting the students’ privacy). The number of examples, attributes and classes of these data sets is shown in Table 4-2.

Table 4-2 Main characteristics of the data sets used in the experiments

Data set # examples # attributes # classes

IRIS flower data set 150 4 3

Multi enrollment data set 460 5 6

38

4.1 Experiment I: IRIS dataset

The IRIS data takes example from two characteristic marks of flowers: sepal and petal.

Researchers can infer a flower’s species from its sepal length, sepal width, petal length, and petal width. Input data is numerical value; output class consists three species: Setosa, Versicolor, and Virginica.

4.1.1 Descriptive statistics

The attributes and class of IRIS data set are listed in Table 4-3. The attributes of IRIS flower data set were previously discretized into discrete values for ant colony optimization and traditional affinity set using k-means cluster, denoted as sl-1, sl-2, sl-3, sw-1, sw-2, etc.

Before experiment, the parameters settings for ACO, affinity set, and proposed multi-objective affinity set are given below.

39

Table 4-3 Attributes and class coding of IRIS dataset Attributes Sepal length Average: 5.843

Standard deviation: 0.825 Sepal width Average: 3.0573

Standard deviation: 0.434 Petal length Average: 3.758

Standard deviation: 1.759 Petal width Average: 1.199

Standard deviation: 0.760 Class Species Setosa, Versicolor, Virginica

40

4.1.2 Experiment Results

Custom dividing point for ACO and Traditional Affinity Set is listed as follows: (k-means)

Table 4-4 Custom dividing point for ACO and Traditional Affinity Set Sepal length Dividing point 1: Sepal width <5.6

Dividing point 2: 5.6<= Sepal width <=6.5 Dividing point 3: Sepal width >6.5

sl-1

sl-2

sl-3

Sepal width Dividing point 1: Sepal width<2.8

Dividing point 2: 2.8<= Sepal width <=3.4 Dividing point 3: Sepal width >3.4

sw-1

sw-2

sw-3

Petal length Dividing point 1: Petal length <3

Dividing point 2: 3<= Petal length<=5.1 Dividing point 3: Petal length >5.1

pl-1

pl-2

pl-3

Petal width Dividing point 1: Petal width<1

Dividing point 2: 1<= Petal width<=1.7 Dividing point 3: Petal width>1.7

pw-1

pw-2

pw-3

41

Table 4-5 Parameters settings for experiment I

Methodology Parameters Value

Ant colony optimization Folds 10

Number of ants 10

Default class Virginica

Minimum cases per rule 5

Maximum uncovered cases 10

Rules for convergence 10

Number of iterations 100

Affinity set Selection of k 32%

Default class Virginica

Multi-objective affinity set Maximum number of rules (N) 5

Default class Virginica

Iteration times 10000

42

The following presents the classification rules below and comparison results in Table 4-9.

Rule set from ant colony optimization:

Table 4-6 Rule set from ant colony optimization Rule 1 IF Petal length = pl-3

THEN Species = Setosa

Rule 2 IF Petal length = pl-1 AND PetalWidth2 = pl-3

THEN Species = Versicolor

Rule 3 IF Petal length = pl-2

THEN Species = Virginica

Rule 4 IF Sepal length = pl-1

THEN Species = Virginica

Default Species = Virginica

43

Rule set from multi-objective affinity set:

Table 4-7 Rule set from multi-objective affinity set

Rule 1 IF Sepal length= sl-2 AND Sepal width= sw-1 AND Petal length = pl-1

AND Petal width = pw-3

THEN Species= Versicolor

Rule 2 IF Sepal width= sw-1 AND Petal width = pw-1

THEN Species= Setosa

Rule 3 IF Petal length = pl-1 AND Petal width = pw-3

THEN Species= Versicolor

Rule 4 IF Petal width = pw-1

THEN Species= Setosa

Default Species = Virginica

From the results in Table 4-6 and Table 4-7, rule set generated by proposed model seems to contain more variety; thus, these rules keep more information and are more meaningful for botanist or biologist to analyze.

44

Table 4-8 Dividing point for multi-objective affinity set Sepal length Dividing point 1: Sepal width <5.62

Dividing point 2: 5.62<= Sepal width <=6.82 Dividing point 3: Sepal width >6.82

sl-1

sl-2

sl-3

Sepal width Dividing point 1: Sepal width <2.88

Dividing point 2: 2.88<= Sepal width <=3.68 Dividing point 3: Sepal width >3.68

sw-1

sw-2

sw-3

Petal length Dividing point 1: Petal length <3.16

Dividing point 2: 3.16<= Petal length <=5.13 Dividing point 3: Petal length >5.13

pl-1

pl-2

pl-3

Petal width Dividing point 1: Petal width <0.98

Dividing point 2: 0.98<= Petal width <=1.78 Dividing point 3: Petal width >1.78

pw-1

pw-2

pw-3

This discretization result shows that the proposed model and k-means split the continuous data with some near split points and same number of split points; which means they generate in similar quality.

45

Table 4-9 Classification results of experiment I

Algorithm Accuracy # rules

Ant colony optimization 76.67% 4

Affinity set 88.00% 6

*Multi-objective affinity set 97.33% 4

* denotes the best model

In this IRIS classification case, the result shows Multi-objective affinity set seems to be the best model among three classification methods. In the next two sections, we applied and compared the three models on practical issues for experiment.

46

4.2 Experiment II: Multiple Enrollment Dataset

The multi enrollment dataset was observed and collected from a 4-year system technical–

vocational college in Taipei in 2001. In the dataset, we choose several attributes such as grade, conduct, sports, etc. to deduce the enrollment type of multi enrollment programs in Taiwan’s technical–vocational college obtained. Since the concept of Multiple Intelligence was proposed by Gardner in 1993 (Gardner, 1993), multi enrollment program became a trend of

enrollment entrances program in many countries. In 1995, Taiwan Ministry of Education proposed ―The report of education in Taiwan, ROC‖ (Ministry-of-Education, 1995), paraded and planned the multi enrollment entrance program to enhance students’ learning effect and

interests by a more adapted selection. Therefore, in this case we experiment the attributes, which can easily be observed and present students’ learning effect, to examine how effectively

multi enrollment worked. The attributes and details are listed in Table 4-10. In addition, students’ personal information was removed for protecting privacy.

47

4.2.1 Descriptive statistics

Table 4-10 Attributes and class coding of IRIS dataset

Attributes Conduct Average: 87.202

Standard deviation: 5.205

Grade Average: 69.741

Standard deviation: 7.503

Sports Average: 76.780

Standard deviation: 10.749 Rank in class Average: 29.407

Standard deviation: 16.804 School Score Average: 87.202

Standard deviation: 5.205 Class Enrollment entrances Application (Application)

Joint entrance examination (Joint) Audition (Audition)

Cerebral palsy disability (Disability) Disaster area admission (Disaster) Technical admission (Tech)

48

4.1.2 Experiment Results

Custom dividing point for ACO and traditional Affinity Set

Table 4-11 Custom dividing point for ACO and traditional Affinity Set Conduct Dividing point 1: Conduct<83

Dividing point 2: 83<= Conduct<=88 Dividing point 3: Conduct>88

Low Middle High Grade Dividing point 1: Grade<65

Dividing point 2: 65<= Grade<=73 Dividing point 3: Grade>73

Low Middle High Sports Dividing point 1: Sports<47

Dividing point 2: 47<= Sports<=76 Dividing point 3: Sports>76

Low Middle High Rank in class Dividing point 1: Rank<21

Dividing point 2: 21<= Rank<=40 Dividing point 3: Rank>40

Front Middle Post School Score Dividing point 1: SchoolScore<83

Dividing point 2: 83<= SchoolScore<=88 Dividing point 3: SchoolScore>88

Low Middle High

49

Table 4-12 Parameters settings for experiment I

Methodology Parameters Value

Ant colony optimization Folds 10

Number of ants 10

Default class Joint

Minimum cases per rule 5

Maximum uncovered cases 10

Rules for convergence 10

Number of iterations 100

Affinity set Selection of k 35%

Default class Joint

Multi-objective affinity set Maximum number of rules (N) 5

Default class Application

Iteration times 10000

50

The following presents the classification rules set in Table 4-13 to 4-15. The comparison result was presented in Table 4-17. Rule set from ant colony optimization:

Table 4-13 Rule set from ant colony optimization Rule 1 IF Conduct = Middle AND Grade = Middle

THEN enrollment= Joint Rule 2 IF Conduct = Low

THEN enrollment= Joint Rule 3 IF ClassRank = Front

THEN enrollment= Joint Rule 4 IF Grade = Low

THEN enrollment= Joint

Rule 5 IF Conduct = High AND Grade = High THEN enrollment=Audition

Rule 6 IF Conduct = High AND Sports = Middle THEN enrollment= Joint

Default enrollment=Joint

51

Rule set from affinity set:

Table 4-14 Rule set from affinity set Rule 1 IF Sports=Middle

THEN Enrollment= Joint Rule 2 IF Grade= Middle

THEN Enrollment= Joint Rule 3 IF SchoolScore= Middle

THEN Enrollment= Joint Rule 4 IF Conduct= Middle

THEN Enrollment= Joint

Rule 5 IF Conduct= Middle AND SchoolScore= Middle THEN Enrollment= Joint

Rule 6 IF Grade= Middle AND Sports= Middle THEN Enrollment= Joint

Rule 7 IF Sports= Middle AND SchoolScore= Middle THEN Enrollment= Joint

Rule 8 IF Conduct= Middle AND Sports= Middle THEN Enrollment= Joint

52

Rule 9 IF Conduct= Middle AND Sports= Middle AND SchoolScore= Middle THEN Enrollment= Joint

Rule 10 IF ClassRank= Middle THEN Enrollment= Joint Default enrollment=Joint

In this multi enrollment case, we selected rules by setting k=35%, and obtained totally 10 rules; however, the k-core method cannot help us to avoid the rule set pointing to the same class, thus the rule set cannot reveal the feature of the dataset. On the contrary, our proposed multi-objective affinity set has the advantage to avoid the kind of issues happen by using iteration selection. Rule set from multi-objective affinity set is listed as follow in Table 4-15.

The proposed multi-objective affinity set output totally 4 rules, and is fewer than ACO (6 rules) and traditional affinity set (10 rules). These four rules highlight the ―School Score‖ ,an

integrated number considered attendance, bonus point by teachers, merits, and faults , might be a major attribute that shows the difference of students’ learning effects by different

enrollment entrance.

53

Table 4-15 Rule set from multi-objective affinity set Rule 1 IF SchoolScore =Low

THEN enrollment= Joint

Rule 2 IF Grade = Low AND SchoolScore =Middle THEN enrollment= Joint

Rule 3 IF Sports = Low AND SchoolScore =High THEN enrollment= Joint

Rule 4 IF Grade = Middle AND Sports = Low THEN enrollment= Joint

Default enrollment=Application

Dividing points is shown in Table 4-16.

Notice that in attribute ―Conduct‖ and ―Rank in class‖ both have only one dividing point, less

than two that k-means generated.

54

Table 4-16 dividing point for multi-objective affinity set Conduct Dividing point 1: Conduct<95.0

Dividing point 2: Conduct>=95.0

Low High Grade Dividing point 1: Grade<49.41

Dividing point 2: 49.41<= Grade<=62.87 Dividing point 3: Grade>62.87

Low Middle High Sports Dividing point 1: Sports<23.83

Dividing point 2: 23.83<= Sports<=57.23 Dividing point 3: Sports>57.23

Low Middle High Rank in class Dividing point 1: Rank<62

Dividing point 3: Rank>=62

Front Post School Score Dividing point 1: SchoolScore<81.17

Dividing point 2: 81.17<= SchoolScore<=87.80 Dividing point 3: SchoolScore<87.80

Low Middle High

55

The following is the comparison of three models.

Table 4-17 Classification results of experiment II

Algorithm Accuracy # rules

Ant colony optimization 61.44% 6

Affinity set 61.30% 8

*Multi-objective affinity set 61.50% 4

* denotes the best model

This result shows that the proposed multi-objective affinity set has advantages to enhance accuracy, and decrease the number of split points and rules. Fewer rules without losing information and variety can be more easily applied and build for education diagnosis system.

56

5. CONCLUSION

5.1 Conclusion

The major purpose of this research is to combine multi-objective decision making and affinity set classification method, and enhance accuracy of output rule set. Since skipping the k-core method of traditional affinity set, the combination of rules has more variety to be chosen and has a higher prediction accuracy of the three experiments in this study. Furthermore, our improved multi-objective affinity set can reduce the necessary numbers of classification rules.

As a result, our method improves the prediction accuracy via fewer classification rules, and makes the system based on classification rules in real world easier to be applied or constructed, such as web interface on internet, educational support software on PC, etc.

5.2 Future Works

This study focuses on increasing classification accuracy and reducing the number of dividing points and number of classification rules. Since skipping the k-core method of traditional affinity set, the combination of rules has more variety to be chosen and has a higher prediction accuracy of delayed diagnosis detection. Moreover, there are still objectives can be added to

57

the MO affinity set system, such as higher TN, lower FP, etc. On future applications, the focus of the improvement of the multi-objective model should aim at real-world problems, such as

making the system more sensitive for predicting some particular attributes. For example, medical diagnosis system via observing patients’ blood pressure, body temperature, pulse, etc.,

could be used to prevent delayed diagnosis or medical error.

58

REFERENCES

1. Ahmad, M. A., & Srivastava, J. (2008). An Ant Colony Optimization Approach to Expert Identification in Social Networks. Social Computing, Behavioral Modeling, and Prediction, 120-128.

2. Barakat, N. H. (2007). Rule Extraction from Support Vector Machines: A Sequential Covering Approach. IEEE Transactions on Knowledge and Data Engineering, 19(6), 729-741.

3. Berrado, A., & Runger, G. C. (2007). Using Metarules To Organize And Group Discovered Association Rules. Data Mining and Knowledge Discovery, 14(3), 409-431.

4. Brauers, W. K. M., Zavadskas, E. K., Peldschus, F., & Turskis, Z. (2008). Multi-objective decision-making for road design. Transport 23(3), 183 - 193.

5. Chen, Y.-W., & Larbani, M. (2007). Affinity Set and Its Applications. Paper presented at the Proceeding of the International Workshop on Multiple Criteria Decision Making, Poland.

6. Chen, Y.-W., Larbani, M., Shen, C.-M., & Chen, C.-W. (2008). Using Affinity Set on Finding the Key Attributes of Delayed Diagnosis. Applied Mathematical Sciences, 3(7), 217-316.

59

7. Chen, Y.-W., Larbani, M., Wu, C.-L., & Chen, C.-W. (2007). Using Affinity Set Theory to Enhance the Effectiveness of Head Computed Tomography.

8. Coello, M. R.-s. C. A. C. (2006). Multi-Objective particle swarm optimizers: A survey of the state-of-the-art. International Journal of Computational Intelligence Research, 2(3), 287-308.

9. Deb, K. (2001). Multi-Objective Optimization using Evolutionary Algorithms: John Wiley

& Sons, England.

10. Dougherty, J., Kohavi, R., & Sahami, M. (1995). Supervised and Unsupervised Discretization of Continuous Features. Paper presented at the Machine Learning:

Proceeding of the Twelve International Conference.

11. Gardner, H. (1993). Multiple intelligences: The theory in practice. New York: Basic Books.

12. Grzymala-Busse, J. W. (2002). Data reduction: discretization of numerical attributes. In Handbook of data mining and knowledge discovery. New York, NY: Oxford University

Press, Inc.

13. Ho, D. Y. F. (1998). Interpersonal Relationships and Relationship Dominance: An Analysis Based on Methodological Relationism. Asian Journal of Social Psychology, 1, 1-16.

60

14. Holden, N., & Freitas, A. A. (2004). Web Page Classification with an Ant Colony Algorithm. Lecture Notes in Computer Science, 3242, 1092-1102.

15. Hwang, K.-K. (1987). Face and Favor: The Chinese Power Game. The American Journal of Sociology, 92(4), 944-974.

16. Ishibuchi, H., Nakashima, T., & Nii, M. (2005). Multi-Objective Design of Linguistic Models. In Classification and Modeling with Linguistic Information Granules (pp.

131-141): Springer Berlin Heidelberg.

17. Jensen, R., & Shen, Q. (2006). Webpage Classification with ACO-enhanced Fuzzy-Rough Feature Selection. Paper presented at the Proceedings of the Fifth International

Conference on Rough Sets and Current Trends in Computing (RSCTC 2006), LNAI 4259.

18. Kennedy, J., & Eberhart, R. C. (1995). Particle swarm optimization. Paper presented at the IEEE Int. Conf. on Neural Networks, Piscataway, NJ.

19. Kerber, R. (1998). Chimerge: Discretization of numeric attributes. Paper presented at the the 10th Conference of the American Association for Artificial Intelligence.

20. Kianmehr, K., Alshalalfa, M., & Alhajj, R. (2008). Effectiveness of Fuzzy Discretization for Class Association Rule-Based Classification. Paper presented at the Foundations of

Intelligent Systems.

21. Lhotská, L., Macaš, M., & Burša, M. (2006). PSO and ACO in Optimization Problems.

61

Paper presented at the Intelligent Data Engineering and Automated Learning – IDEAL 2006.

22. Liu, H., Hussain, F., Tan, C. L., & Dash, M. (2002). Discretization: An Enabling Technique. Data Mining and Knowledge Discovery, 6(4), 393-423.

23. Luo, Y. (2000). Guanxi and Business (Vol. 1): World Scientific.

24. Mendelson, & B. (1990). Introduction to Topology. Dover Publications.

25. Ministry-of-Education. (1995). An Report of education in Taiwan, ROC o. Document Number)

26. Mostaghim, S. (2003). The Role of -dominance in Multi Objective Particle Swarm Optimization Methods. Paper presented at the Proceedings of the 2003 Congress on

Evolutionary Computation.

27. Pal, P. K. T. S. B. S. K. (2007). Multi-Objective Particle Swarm Optimization with time variant inertia and acceleration coefficients Information Sciences, 177(22), 5033-5049 28. Pfahringer, B. (1995). Compression-Based Discretization of Continuous Attributes. Paper

presented at the Proceedings of the 12th International Conference on Machine Learning.

29. Piatrik, T., & Izquierdo, E. (2006). Image Classification Using an Ant Colony Optimization Approach. Lecture Notes in Computer Science, 4306, 159-168.

30. Qu, W., Yan, D., Sang, Y., Liang, H., Kitsuregawa, M., & Li, K. (2008). A Novel Chi2

62

Algorithm for Discretization of Continuous Attributes. In Progress in WWW Research and Development (Vol. 4976): Springer-Verlag Berlin Heidelberg.

31. Skubacz, M., & Hollmén, J. (2008). Quantization of Continuous Input Variables for Binary Classification. Paper presented at the Intelligent Data Engineering and Automated

Learning — IDEAL 2000. Data Mining, Financial Engineering, and Intelligent Agents.

32. Wang, Z., Sun, X., & Zhang, D. (2007). A PSO-Based Classification Rule Mining Algorithm (Vol. 4682). Heidelberg: Springer Berlin.

33. Wu, C.-H., Lin, W.-T., Li, C.-H., Fang, I.-C., & Wu, C.-H. (2008). Ant Colony Optimization On Building An Online Delayed Diagnosis Detection Support System For

Emergency Department. Paper presented at the CIEF 2008.

34. Wu, C.-H., Lin, W.-T., Li, C.-H., Fang, I.-C., & Wu, C.-H. (2009). A Novel Multi-Objective Affinity Set Classification System: An Investigation of Delayed Diagnosis

Detection. Paper presented at the 1st Asian Conference on Intelligent Information and

Database Systems.

相關文件